JP2024513845A - System and method for automatic template-based detection of anatomical structures - Google Patents

Abstract

The system and method for detecting anatomical structures includes generating an initial anatomical template for each training subject of a plurality of training subjects based on a corresponding MR image and a first training subject of the plurality of training subjects. The computing device can map the MR images of the other training subjects onto the template space by applying a global transformation followed by a local transformation, can average the mapped MR images with the initial anatomical template to generate a final anatomical template, and can draw a boundary of the anatomical structure of interest on the final anatomical template. The computing device can refine the boundary using an edge detection algorithm. The final anatomical template can be used to automatically (e.g., without human intervention) identify the boundary of the anatomical structure of interest in the untrained subjects.

Description

CROSS-REFERENCE TO RELATED APPLICATIONS This application is filed in U.S. Provisional Patent Application No. 63/170, 2, filed on April 2, 2021, and entitled “SYSTEMS AND METHODS FOR AUTOMATIC TEMPLATE-BASED DETECTION OF ANATOMICAL STRUCTURES.” No. 29 priority claims, the contents of which are incorporated herein by reference in their entirety.

TECHNICAL FIELD This disclosure relates generally to the field of detecting or identifying boundaries of anatomical structures. In particular, the present disclosure relates to methods and systems for generating standardized templates and using such templates to automatically detect or identify boundaries of anatomical structures.

One disease that requires identification of anatomical boundaries is Parkinson's disease (PD), a chronic progressive neurodegenerative disorder that affects approximately 1% of individuals over the age of 60. PD is pathologically characterized by early neurodegeneration of the substantia nigra (NM) in the substantia nigra pars compacta (SNpc) and increased iron deposition in the substantia nigra (SN). Degeneration of the SN is a hallmark of the progression of numerous neurodegenerative diseases. Widespread neuronal loss of the SNpc also occurs in atypical parkinsonian disorders, including progressive supranuclear palsy (PSP) and multiple system atrophy (MSA), in which different subregions of the SN are affected.

The SN is composed of two anatomically and functionally distinct regions, the SN pars reticularis (SNpr) and the SNpc. The SNpc contains a dense distribution of NM containing dopaminergic neurons, while iron content tends to be higher in the SNpr. However, clusters of SNpc dopaminergic neurons (known as nigrosomes) are deeply embedded within the SNpr, and therefore the boundary between the SNpr and SNpc is difficult to delineate, especially in the caudal region of the SN. It is. The regional selectivity of PD is relatively specific with a 50%-70% loss of pigmented neurons in the ventrolateral layer of the SNpc (at the time of symptom onset). Regarding the SN, iron deposition and volume changes in the red nucleus (RN) and subthalamic nucleus (STN) have been reported to be associated with disease status and progression rate, and deep brain stimulation (DBS) treatment in PD patients has also been reported. serves as an important target for

According to at least one aspect, a magnetic resonance imaging (MRI) system includes: an MRI scanner configured to acquire magnetic resonance (MR) data; at least one processor; a memory in which code instructions are stored. The computer code instructions, when executed by the at least one processor, cause the at least one processor to instantiate one or more anatomical structures of interest through an MRI scanner for each training subject of the plurality of training subjects. A corresponding MR image with a contrast of At least one processor can generate an initial anatomical template image based on first MR data of a first training subject of the plurality of training subjects. The initial anatomical template image may define a template space. For each training subject of the plurality of training subjects other than the first training subject, at least one processor (i) applies a global transform to the MR images of the training subject to (ii) applying a local transformation to the first morphing version of the training subject's MR image to create a first morphed version of the training subject's MR data in template space; A second morphed version of the MR image representing the second estimate can be generated. At least one processor can average the initial anatomical template image and the second morphed version of the MR images of the plurality of training subjects to generate a final anatomical template image. At least one processor delineates boundaries of one or more anatomical structures of interest in a final anatomical template, and uses the final anatomical template to delineate boundaries of one or more anatomical structures of other non-training subjects. Boundaries of anatomical structures of interest can be identified.

According to at least one aspect, a method acquires, via an MRI scanner, a corresponding MR image with contrast illustrating one or more anatomical structures of interest for each training subject of the plurality of training subjects. This may include: The method can include the computing device generating an initial anatomical template image based on first MR data of a first training subject of the plurality of training subjects. The initial anatomical template image may define a template space. For each training subject of the plurality of training subjects other than the first training subject, the computing device (i) applies a global transform to the MR image of the training subject to create a first set of MR data of the training subject in template space; (ii) applying a local transformation to the first morphing version of the training subject's MR image to represent the first morphed version of the training subject's MR data in template space; A second morphed version of the MR image representing an estimate of 2 can be generated. The method may include averaging the initial anatomical template image with a second morphed version of MR images of a plurality of training subjects other than the first training subject to generate a final anatomical template image. can. The method includes delineating the boundaries of one or more anatomical structures of interest in a final anatomical template and using the final anatomical template to delineate one or more anatomical structures of interest in other non-trained subjects. identifying the boundaries of the anatomical structure;

According to at least one aspect, non-transitory computer-readable media can include computer code instructions stored on the non-transitory computer-readable media. The computer code instructions, when executed by the processor, cause the processor to scan a corresponding MR image with contrast illustrating one or more anatomical structures of interest through an MRI scanner for each training subject of the plurality of training subjects. Images can be acquired. The processor can generate an initial anatomical template image based on the first MR data of a first training subject of the plurality of training subjects. The initial anatomical template image may define a template space. For each training subject of the plurality of training subjects other than the first training subject, the processor (i) applies a global transform to the MR images of the training subject to obtain a first estimate of the training subject's MR data in template space; (ii) applying a local transformation to the first morphed version of the training subject's MR image to create a second morphed version of the training subject's MR data in template space; A second morphed version of the MR image representing the estimate can be generated. The processor may average the initial anatomical template image with a second morphed version of MR images of a plurality of training subjects other than the first training subject to generate a final anatomical template image. The processor delineates the boundaries of the one or more anatomical structures of interest in the final anatomical template and uses the final anatomical template to delineate the one or more anatomical structures of interest in other non-training subjects. Be able to identify the boundaries of a scientific structure.

3 is a flowchart illustrating a method for generating an anatomical template in accordance with the inventive concepts of the present disclosure. 2 is a flowchart illustrating a method for automatic template-based identification of anatomical structure characteristics in accordance with the inventive concepts of the present disclosure. 1 shows a diagram illustrating a system for detecting or identifying boundaries of an anatomical structure in accordance with the inventive concepts of the present disclosure; FIG. Magnetic resonance (MR) images depicting the NM perspective in NM-MRI template space and the SN, RN, and STN perspectives in QSM template space are shown. Figure 3 shows images illustrating various steps for mapping NM boundaries from template space to original space. Figure 3 shows images illustrating various steps for mapping SN, RN boundaries and STN boundaries from template space to original space. Figure 3 shows Dice similarity coefficients plotted against neuromelanin (NM) volume ratio values based on various scenarios. Figure 3 shows Dice similarity coefficients plotted against neuromelanin (NM) volume ratio values based on various scenarios. Figure 3 shows Dice similarity coefficients plotted against substantia nigra (SN) volume ratio values based on various scenarios. Figure 3 shows Dice similarity coefficients plotted against substantia nigra (SN) volume ratio values based on various scenarios. 1 shows the Dice similarity coefficient plotted against the volume ratio values of the red nucleus (RN) under various scenarios. Figure 3 shows the Dice similarity coefficient plotted against subthalamic nucleus (STN) volume ratio values based on various scenarios. Shows plots illustrating the correlation between SN, RN, and STN iron content resulting from manual and template segmentation of 30 training cases (top plot) and 57 validation cases (bottom plot). . A plot depicting the agreement between manual and template neuromelanin (NM) background measurements for 30 healthy controls is shown. FIG. 3 shows an image illustrating background boundaries relative to NM boundaries identified using the Dynamic Programming Algorithm (DPA). An image is shown illustrating the background boundary and the NM DPA boundary after transformation from template space to original space. Figure 3 shows images illustrating boundaries reconstructed with various values of the DPA parameter α. We present simulation results for using DPA to identify the boundary of a rectangular region with two different intensities. We show simulation results for using DPA to identify the boundary of a crescent-shaped region with two different intensities. We show simulation results for using DPA to identify the boundaries of cashew-shaped regions with two different intensities.

For some diseases, such as Parkinson's disease (PD), information about the disease state and/or useful in some treatments can be estimated from the volume or other geometric properties of some anatomical structures. can be done. In PD, iron deposition and volume changes in the substantia nigra (SN), red nucleus (RN) and subthalamic nucleus (STN) are known to be associated with disease status and progression rate; It serves as an important target for deep brain stimulation (DBS) therapy in PD patients. Therefore, an accurate and comprehensive in vivo delineation of the SN and SN subregions, RN, and STN is essential to detect changes in iron and neuromelanin (NM) composition in PD and other movement disorders affecting the midbrain. It may be useful to investigate fully. Therefore, accurate in vivo delineation of the SN, subregions of the SN, and other midbrain structures such as the red nucleus (RN) and subthalamic nucleus (STN) is essential to fully investigate iron and NM changes in PD. could be useful.

To date, many studies still use manual or semi-automated approaches to define deep brain gray matter. However, manual segmentation is time consuming, especially when large amounts of data need to be evaluated. Also, unless the rater is well trained, manual drawings have low reliability of reproducibility between individuals or between sites. Some approaches for in vivo delineation can include the use of templates to map iron and NM content. In such an approach, creating a standardized template can (a) recognize changes in iron and NM distributions; (b) automatically calculate volumes associated with such distributions; (c) quantifying iron content and NM signal changes; and (d) can significantly impact the reliability of these measurements. Anatomical templates of the SN using conventional structural MR sequences have been used previously in several studies. T1-weighted (T1W) and T2-weighted (T2W) images can be employed to create a high-resolution probabilistic in vivo subcortical nucleus atlas, which distinguishes the SN from the substantia nigra pars compacta (SNpc) and the substantia nigra. It is now possible to segment the pars reticularis (SNpr). Such atlases may not be suitable for studies in older adults if they are based on images from a young adult population (e.g. mean ± standard deviation age: 28.9 ± 3.6 years). . Furthermore, both T1W and T2W contrast images cannot be easily used to delineate the highly interdigitated boundary between SNpc and SNpr in elderly subjects. Based on the anatomical connectivity of SN subregions to different parts of the brain, we use diffusion-based tractography to segment the SN into SNpc and SNpr, or the SN/ventral tegmental area (VTA). can be subdivided into dorsomedial and ventrolateral subregions. However, the estimated structural connectivity has known biases and is highly dependent on the data acquisition and fiber tracking algorithms, and diffusion imaging suffers from low resolution data acquisition. Therefore, direct visualization and segmentation of the SN and subregions of the SN using high-resolution imaging would be a more accurate option, especially when it comes to detecting subtle pathological changes in the SN.

One approach to determine the boundaries of mesencephalic nuclei and examine pathological changes in PD patients is T2 ^* -weighted gradient echo (GRE) imaging and susceptibility-weighted imaging (GRE), where iron-containing regions appear hypointense. SWI). On the other hand, the development of quantitative susceptibility mapping (QSM) allows the quantification of iron stored in ferritin and hemosiderin. QSM-based techniques have shown that tissue magnetic susceptibility correlates well with brain iron in PD patients. In addition, QSM is superior to traditional T2 ^* W imaging for preoperative target guidance of DBS. However, QSM alone cannot separate SNpc from SNpr because both SNpc and SNpr contain iron. This limitation can be overcome by using neuromelanin-sensitive MRI (NM-MRI), which has been developed in the past few years. The SNpc and ventral tegmental area (VTA) are primarily composed of NM-containing dopaminergic neurons, whereas the SNpr is not. Therefore, the hyperintense signals seen on NM-MRI in the midbrain are spatially associated with regions of the SNpc and VTA (this has been verified by post-mortem histological studies). Therefore, the overlap between the volume of NM (SNpc plus VTA) and the volume of iron-bearing SN (SNpc plus SNpr) is considered to represent SNpc.

So far, existing approaches do not accurately and reliably segment SNpc using QSM and/or NM-MRI imaging. The SN atlas can be obtained using QSM only. Alternatively, NM templates based on NM-MRI imaging may be employed, for example created using manual drawing, automatic segmentation, or artificial intelligence. The purpose of using templates would be to facilitate the identification of boundaries. However, template mapping is not perfect, and the ideal space for marking boundaries is in the original, pristine data space. A simple threshold, whether in template space (a standard fixed brain or volunteer, to which all other brains are mapped) or in original space (original brain imaging data for a given individual) There are drawbacks to this method. Setting the threshold to a relatively high or relatively low value can dramatically reduce the estimated amount of NM or iron content, especially in PD patients with severe NM degeneration and iron deposition in the SN. It could lead to change. Also, the variable contrast of the image can make it difficult to use some algorithms such as region growing. Even though the SN has a high iron content compared to surrounding areas, the iron is not distributed uniformly and there is a reduction in iron in the nigrosome 1 (N1) territory. Gaps in structures such as the N1 territory can also increase the difficulty of automatically segmenting SN and NM regions of interest (ROIs).

In the current disclosure, a system and method for detecting anatomical structures can include using high-resolution imaging from a single multi-echo NM MRI sequence to create both NM and iron templates and calculate the boundaries of each structure in template space. Specifically, both NM and iron templates from a single high-resolution NM MRI sequence can be used to (i) calculate the boundaries of each structure in template space, (ii) map these boundaries back to the original space, and then (iii) use a dynamic programming algorithm (DPA) to fine-tune the boundaries in the original space to match the details of each individual's NM and SN characteristics. Experimental results for SN NM and SN, STN, and RN iron content all show strong agreement in terms of Dice values and volume ratios, with the former being 0.85, 0.87, 0.75, and 0.92, respectively, and the latter being 0.99, 0.95, 0.89, and 1.05, respectively. These high quality results demonstrate that it is possible to measure tissue attributes such as volume, iron content, and neuromelanin content with accuracy. Multiple sequences sensitive to NM and iron may also be used, in which case co-registration between the two may be required.

Referring to FIG. 1, a flowchart is shown illustrating a method 100 for generating an anatomical template in accordance with the inventive concepts of the present disclosure. The method 100 includes: a computing device or system acquiring a corresponding MR image for each training subject of a plurality of training subjects (step 102); generating an initial anatomical template based on the acquired MR images (step 104). Method 100 includes a computing device or system mapping MR images of other training subjects onto a template space defined by an initial anatomical template by applying a global transformation followed by a local transformation. (Step 106). Method 100 may include averaging the mapped MR image and the initial anatomical template to generate a final anatomical template (step 108). The method 100 may include drawing boundaries of the anatomical structure of interest in the final anatomical template (step 110). Method 100 may include using the final anatomical template to identify boundaries of anatomical structures of interest in a non-training subject (step 112).

Method 100 can include a computing device or system acquiring a corresponding MR image for each training subject of a plurality of training subjects (step 102). Training subjects may be selected based on some predefined criteria, such as age, health status, or each subject's medical condition and/or medical history. The study described herein was approved by the local ethics committee, and all subjects signed a consent form. Eighty-seven healthy subjects (mean age: 63.4 ± 6.2 years, range: 45-81 years, 53 women) were recruited from the community by advertisement. Exclusion criteria for potential training subjects were (a) structural abnormalities such as tumors, subdural hematomas, or contusions from previous head trauma; (b) history of stroke, addiction, neurological or psychiatric disease; (c) included one or more macrovascular disease diseases with large volume white matter lesions (e.g., Fazekas grade III);

MR images of the training subject may be acquired via an MRI scanner. For example, in a study conducted, MR imaging was performed on a 3T Ingenia scanner (Philips Healthcare, Netherlands) using a 15-channel head array coil. Imaging parameters for the 3D gradient echo SWI sequence with actuated magnetization transfer contrast (MTC) pulses were: time to echo (TE) = 7.5 ms, ΔTE = 7.5 ms with a total of 7 echoes, repetition time (TR) = 62 ms, flip angle = 30°, pixel bandwidth = 174 Hz/pixel, matrix size = 384 x 144, slice thickness = 2 mm, number of slices = 64, spatial in-plane resolution interpolated to 0.67 x 1.34 mm ² =0.67×0.67 mm ² , including a sensing factor of 2, elliptical sampling of k-space, and a total scan time equal to 4 minutes and 47 seconds. The magnetization transfer (MT) resonance state radiofrequency pulse used a nominal flip angle of 90°, zero frequency offset, and a set of three block pulses, each 1.914 milliseconds (ms) in duration. The minimum acceptable TR was used based on safety considerations for specific absorption rates. Due to this long repetition time of 62 ms, 7 echoes were collected.

In some implementations, the computing device or system uses the first echo of the MTC-SWI magnitude image (TE = 7.5 ms) to determine the NM content because the echo provides the primary MT contrast. can be depicted. The computing device or system may assess iron deposition in the SN using a second echo (TE = 15 ms) or a combination of QSM images from two or more echoes for QSM reconstruction. can. The computing device or system (i) segments the brain using a brain extraction tool, BET, and (ii) unrolls the original topological data using a 3D phase unrolling algorithm (3DSRNCP). , (iii) using Advanced Harmonic Artifact Reduction (SHARP) to remove unwanted background fields, and (iv) cutting k-space partitioning using an iterative approach to reconstruct the final QSM map. A susceptibility map can be created by using a (TKD) based inverse filtering technique.

Manual region of interest (ROI) segmentation can be employed in tracing the boundaries of the anatomy of interest. To measure NM and iron content, regions of interest (ROI) of NM, SN, RN, and STN were quadrupled using SPIN software (SpinTech, Inc., Bingham Farms, MI, USA). The magnitude of the MTC scaled up and can be manually traced by a single evaluator on the QSM map. Trace the NM-based SN boundaries starting from the last caudal slice of 4-5 slices until the NM is no longer visible, when 2 mm thick slices are used (or over a total thickness of 8-10 mm). can do. Iron-based SN boundaries can be traced starting from one slice below the most cranial slice where the subthalamic nucleus is visible, up to the most caudal slice, or over a total thickness of 8 to 12 mm, 4 to 6 It is possible to continue with several consecutive slices. The RN ROI can be outlined starting from the last caudal slice and continued cranially or over 3-4 slices over a total thickness of 6-8 mm. The STN ROI can be traced cranially for two slices or over a total thickness of approximately 4 mm. For all ROIs, a dynamic programming algorithm (DPA) can be used to determine the final boundaries to reduce subjective bias. All these boundaries can then be reviewed one by one by a second reviewer and modified accordingly in agreement with the first reviewer.

It is to be understood that the data acquisition parameters and/or data processing scenarios described above represent example implementations that are provided for purposes of illustration and not limitation. For example, the slice numbers provided above are specific to a given resolution with a slice thickness equal to 2 mm. If we use 1 mm thick slices, all of these slice numbers will be doubled. Additionally, although the description focuses primarily on applications to midbrain PD and anatomy, the methods and techniques described herein have other uses, or include deep cerebellar or striatal nuclei. It may also be applied to other anatomical regions, such as the caudate nucleus, or any other region of interest.

In some implementations, the computing device or system provides an MR scanner with a 7.5 ms echo time for each training subject of the plurality of training subjects for MR data acquisition, e.g., starting from the original whole brain 64 slice data. MTC data acquisition can be performed. The MR scanner can construct a corresponding three-dimensional (3D) image (or a plurality of corresponding two-dimensional (2D) images) for each training subject based on the corresponding acquired MR data. In acquiring MR data of multiple training subjects, the MR scanner is used to perform MR scanning with/with contrast that exemplifies one or more anatomical structures of interest (e.g., regions of the SN, RN, and/or STN). The image can be configured with selected parameters to generate it. The MR scanner can acquire NM-MRI data (eg, NM-MRI images) and/or QSM data (QSM images) for each training subject of a plurality of training subjects.

Method 100 can include a computing device or system generating an initial anatomical template based on a selected MR image of a first training subject of the plurality of training subjects (step 104). . A computing device or system (or a user thereof) may select a first 3D MR image (or a corresponding 2D image) of a first training subject from among the MR images of a plurality of training subjects. The selection can be random or according to predefined preferences or criteria. In generating the initial anatomical template, the computing device or system calculates the The first 3D image (or the corresponding 2D image) can be expanded in a plane by a factor. This step may also include interpolation of the slice selection direction to obtain isotropic resolution in template space. The template space can be defined at a desired resolution or at a resolution higher than that of the acquired MR data. In some implementations, the computing device or system can generate an initial NM-MRI template and an initial QSM template based on the NM data and QSM data of the first training subject, respectively. As discussed in more detail below, a computing device or system can use the initial anatomical template to map MR images of other training subjects into template space.

Method 100 includes a computing device or system mapping MR images of other training subjects onto a template space defined by an initial anatomical template by applying a global transformation followed by a local transformation. step 106). The computing device or system may apply global and local transformations to each of the images of other training subjects (other than the first training subject). The computing device or system magnifies each MR image of the MR images in a plane by a magnification factor (e.g., a factor of 2, 4 or more) for all slices, and then expands each MR image of the MR images of other training subjects other than the first training subject. A global transform can be applied over a series of slices covering a region of interest within each MR image. For example, for the midbrain, if the slice thickness is 2 mm, the series of slices covering the region of interest may be a set of 50 central slices. The global transformation may include a rigid transformation followed by an affine transformation, for example by applying b-spline interpolation using the Insight Segmentation and Registration Toolkit freeware (ITK). The computing device or system can apply the same global transformation to map the QSM data for each of the other training subjects to the initial QSM template.

The computing device or system applies a global transform to each MR image (e.g., QSM image and/or NM-MRI image) of the other training subjects to obtain an MR image representing a corresponding firs estimate of the initial template in template space. A corresponding first morphed version of the image can be generated. That is, a global transform is used to match each MR image of another training subject (other than the first training subject) to the initial anatomical template in template space. However, global transforms typically do not provide an exact match to the anatomical structures in the initial anatomical template.

In order to improve the matching between the transformed MR images and the initial anatomical template in template space, the computing device or system uses the first of the MR images of other training subjects (other than the first training subject). Local transformations can be applied to morphed versions of . The computing device or system crops the first morphed version of the MR image in a plane to cover the midbrain and into approximately 16 slices (e.g., if the slice thickness is 2 mm); Coverage of the midbrain territory can be ensured. The computing device or system can apply a local transformation to the cropped image volume to generate a second morphed version of the MR image of the other training subject. For each MR image of a training subject (other than the training subject), the corresponding second morphed version represents a better match with the initial anatomical template in template space.

As an example implementation, a total of 26 training subjects may be used. A computing device or system may select MR data of one of these training subjects (referred to herein as a first training subject) to generate an initial anatomical template. The computing device or system then applies a global transform and a local transform to the MR data of each of the other 25 training subjects, and after applying the global transform and local transform, the MR data of each of the other 25 training subjects. Each MR data can be mapped to an initial anatomical template in template space.

Method 100 may include averaging the mapped MR image and the initial anatomical template to generate a final anatomical template (step 108). A computing device or system can average the mapped MR images with the initial anatomical template to generate a final anatomical template. The result is an averaged template defined for 16 slices encompassing, for example, midbrain territories. The averaged anatomical template is generated by comparing the initial anatomical template generated based on MR data of a single training subject or by averaging the first morphed version and the initial anatomical template. has more representative boundaries of the anatomical structure of interest compared to the template. The computing device or system may linearly interpolate the averaged anatomical template in the slice selection direction to create a template with an isotropic resolution of, for example, 0.167 (192 slices total).

The method 100 may include drawing boundaries of the anatomical structure of interest on the final anatomical template (step 110). Tracing the boundaries of the anatomical structure (or region) of interest can be performed manually, for example by a doctor or radiologist. The final anatomical template generated according to steps 102-108 of method 100 exhibits enhanced contrast of the anatomical structures of interest, thus allowing more accurate tracing of the boundaries of such structures. .

In the final anatomy template, the average value can be considered as representing a probability map for finding boundaries. In some implementations, more than one final anatomy template may be generated. For example, one QSM template and one NM-MRI template can be generated based on NM data and QSM data, respectively. In some implementations, the boundaries of the NM, red nucleus (RN), SN, and subthalamic nucleus (STN) may all be drawn manually. For neuromelanin (NM) data, slices 44-98 from a total of 192 interpolated slices can be used to draw boundaries, while QSM data slices 44-126 can be used. The actual selection of slice numbers depends on the resolution and degree of interpolation used.

Method 100 may further include fine-tuning the boundaries of the anatomical structure of interest in the final anatomical template using a boundary detection algorithm. The boundary detection algorithm can be implemented as a dynamic programming algorithm (DPA). A computing device or system can run DPA for boundary detection to determine template boundaries. DPA may use a cost function that depends on the local radius of curvature and signal slope. Further details of this algorithm are provided below.

Method 100 may include a computing device or system using the final anatomical template to identify boundaries of anatomical structures of interest in a non-training subject (step 112). The use of the final anatomical template to identify the boundaries of anatomical structures of interest in untrained subjects is discussed in detail with respect to FIG. 2 below.

Referring now to FIG. 2, a flowchart is shown illustrating a method 200 of automatic template-based identification of anatomical structure characteristics in accordance with the inventive concepts of the present disclosure. Method 200 can include acquiring MR images (or MR data) of a non-trained subject (step 202). Method 200 may include mapping MR images of a non-training subject to a final anatomical template in template space (step 204) by applying a global transformation followed by a local transformation. Method 200 may include projecting boundaries of one or more anatomical structures of interest from the final anatomical template onto the mapped MR image of the untrained subject (step 206). Method 200 may also include applying an inverse transformation to the projected boundaries of the anatomy of interest to generate an estimate of the boundaries in the MR image of the untrained subject (step 208).

Method 200 enables automatic identification of boundaries of anatomical structures of interest for untrained subjects using the final anatomical template. Although method 100 for generating anatomical templates may involve human intervention (e.g., manually drawing boundaries in step 110), method 200 for automatic identification of boundaries of anatomical structures of interest for untrained subjects may be performed/executed automatically without any human intervention.

Method 200 may include a computing device acquiring MR images (or MR data) of an untrained subject (step 202). The MR scanner may acquire MR data (e.g., a 3D MR image or multiple 2D MR images) of the untrained subject in a manner similar to that discussed above with respect to step 102 of method 100 of FIG. 1 . The untrained subject may be one (e.g., a patient or a research subject) that is not one of the multiple training subjects used in method 100 to generate the final anatomical template. The MR scanner may acquire NM data and/or QSM data of the untrained subject.

Method 200 may include mapping MR images of a non-training subject to a final anatomical template in template space (step 204) by applying a global transformation followed by a local transformation. The global and local transformations may be similar to those described with respect to step 106 of method 100 of FIG. The computing device or system may first apply a global transform to the MR images (e.g., NM images and QSM images) of the non-training subject to generate a first morphed version of the MR images of the non-training subject. .

After applying the global transform, the computing device or system may crop the first morphed version of the MR image of the non-training subject. For example, a computing device or system can label the RNs in the two central slices of template data and return them to the original whole brain. If the RN is still only evident in two slices, the lowest slice is set to slice 10. If the RN is in three slices, the middle slice is set to slice 10. This provides a means to best center the data before the final transformation back to template space. The computing device or system scales the cropped image volume by a factor of 4 (or other scaling factor) in the plane and then applies a local transformation to the scaled cropped image volume to make it isotropic. A second morphed version of the MR image of the untrained subject can be acquired with the image (transformation, together with interpolation of the slice selection direction, allows the image to be made isotropic). The second morphed version of the MR image of the untrained subject represents a relatively good match of the final anatomical template in template space. The interpolated in-plane resolution provides a more accurate means for estimating the final object size using the DPA algorithm.

Method 200 may include projecting boundaries of one or more anatomical structures of interest from the final anatomical template onto the mapped MR image of the untrained subject (step 206). Once the MR images of the non-training subject are matched or mapped to the final anatomical template (by applying global and local transformations), the computing device or system draws the final anatomical template. Boundaries of the anatomy of interest can be projected onto a second morphed version of the MR image of the untrained subject.

Method 200 may include applying an inverse transformation to the projected boundaries of the anatomy of interest to generate an estimate of the boundaries in the MR image of the untrained subject (step 208). An inverse transformation is performed to map the template boundaries of the anatomy of interest back onto the midbrain (eg, onto the original image space) of the MR image of the untrained subject. However, since everyone is different and template mapping is not perfect, there is no guarantee that the projected boundaries will fit well.

The method 200 further includes the computing device or system using a boundary detection algorithm (or DPA) to fine-tune the boundaries of an anatomical structure of interest in an originally acquired MR image of a non-training subject. can be included. The computing device or system can apply a threshold to the region inside the transformed boundary for the QSM data to remove negative values and determine the corresponding centerline using image thinning. I can do it. To best choose an initial starting point for DPA, both Otsu histogram analysis and a threshold-based approach can be used to determine whether the original boundaries extend too far outside the structure of interest. . For the NM data, the computing device or system uses 25 additional training subjects (or training subjects other than the first A constant equal to four times the background standard deviation averaged over all subjects) can be determined. For QSM data, the starting threshold can be set to zero. When the Otsu threshold gives a value that causes pixels inside the structure (or region) of interest to be removed, the template transformed boundary can be reduced accordingly. For QSM data, if the Otsu threshold exceeds 30 ppb, the threshold can be set to 30 ppb. Finally, the computing device or system can again modify the resulting boundaries using the same DPA used in the template space (as discussed above with respect to FIG. 1). This DPA, together with RN boundaries, prevents leakage of SN into RNs and with the help of template boundaries, STNs can be distinguished from SNs.

Referring to FIG. 3, a diagram illustrating a system 300 for detecting or identifying boundaries of an anatomical structure in accordance with the inventive concepts of the present disclosure. System 300 can include an MR scanner 302 for acquiring MR data and a computing device 304 for processing the acquired MR data to detect or identify boundaries of anatomical structures. . Computing device 304 may include a processor 306 and memory 308. Memory 308 can store computer code instructions for execution by processor 306. The computer code instructions, when executed by processor 306, can cause computing device 304 to perform method 100 and/or method 200 discussed above. Computing device 304 may further include a display device 310 for displaying MR images and outputting results of method 100 and/or method 200, or other data.

In some implementations, computing device 304 may be an electronic device separate from MR scanner 302. In such implementations, computing device 304 may or may not be communicatively coupled to MR scanner 302. For example, MR data acquired by MR scanner 302 may be transferred to computing device 304 via flash memory, compact disc (CD), or other storage device. When computing device 304 is communicatively coupled to MR scanner 302, MR data acquired by MR scanner 302 is transmitted to computing device 304 via any communication link between MR scanner 302 and computing device 304. may be transferred to device 304.

In some implementations, computing device 304 can be integrated within MR scanner 302. For example, processor 306, memory 308, and/or display device 310 may be integrated within MR scanner 302. In such implementations, MR scanner 302 can acquire MR data and perform method 100 and/or method 200.

In summary, for each of the SN, RN, and STN, boundaries can be initially drawn manually in template space, and system 300 or computing device 304 can run a DPA to fine-tune the boundaries. can. System 300 or computing device 304 can map these boundaries back to the original space and run the DPA again to provide the final boundaries, making this a fully automated process. After the manual drawing is created, the system 300 or computing device 304 can run a DPA that fully automates the final boundary determination. Using the final boundaries, system 300 or computing device 304 can calculate volume, signal strength, sensitivity, and/or overlapping fraction.

The system 300 or computing device 304 calculates the NM complex volume (SNpc plus VTA) and the SN volume (SNpc and SNpr) by overlaying two ROIs from the MTC data and QSM data, respectively. The overlap between the iron content and the iron content can be calculated. The system 300 or computing device 304 uses the iron-containing SN volume to create an overall overlapping fraction (OOF) measure, which is essentially a measure of the SNpc fraction over the entire SN volume, as shown below. Duplicates can be normalized.

The process of generating the final anatomical template described with respect to Figure 1 above was evaluated using MR data from various control subjects. The evaluation was based on a comparison of the performance of the process to generate the final anatomical template against manually drawn boundaries of the anatomical structure of interest.

A total of 87 healthy controls (HC) were scanned. SN, STN, and RN were manually traced for all 87 cases. Of these, 30 individuals (test dataset: including 17 males, 13 females, age range 66±7.2 years) were used for initial training of the template approach described above. Once all aspects of the algorithm were in place, we then validated the method on the following 57 cases (validation dataset: including 17 males and 40 females, age range 61.9±5.0 years). . Two measures were used to evaluate the performance of the anatomical template generation process. These are dice showing the spatial overlap between the structures associated with the manual and template segmentation methods and the volume ratio (VR) of the structures from the template volume divided by the volume of the manual segmentation. It is a similarity coefficient.

Finally, all data were combined to generate quantitative information on structural volume, NM and iron content, and overlap between SN and NM regions. Total iron content was calculated by summing the product of the volume and the average sensitivity of the structure over all slices in which the structure was drawn. Similarly, total NM content was resulting from the summation of the product of NM volume and NM contrast over the corresponding slice.

Referring to FIG. 4, an MR image is shown showing the NM viewpoint in the NM-MRI template space and the SN, RN, and STN viewpoints in the QSM template space. Specifically, images (a)-(c) illustrate various views of the NM 402 in the NM-MRI template using the validation dataset. The boundaries of NM402 are drawn manually. Images (d)-(f) illustrate the viewpoints of SN 404, RN 406, and STN 408. The boundaries of SN 404, RN 406, and STN 408 in images (d)-(f) of Figure 4 were drawn manually.

Referring to FIG. 5, images are shown illustrating various steps for mapping NM boundaries from template space to original space. Specifically, boundaries from 16 slices taken every 12th slice in 0.167 mm isotropic template space are superimposed on the original 2 mm thick midbrain slice and the NM data shown for. Since the midbrain is only visible on these four slices, the border appears only on four of these slices. Each column of images in FIG. 5 represents a separate slice. The top row images, designated as row A, show images of various slices of the constructed neuromelanin template. The second row (from the top), designated as row B, shows the same image as row A, but with a neuromelanin DPA border 502. The third row, designated as row C, shows a superimposed template border 504 representing the transformed border 502 on the original image 504. The final row, designated as row D, shows the subject's midbrain image with the final boundary 506 of the NM obtained by fine-tuning the superimposed template boundary 504 using DPA.

Referring to FIG. 6, which shows images illustrating various steps for mapping the SN, RN, and STN boundaries from template space to the original image space for QSM data, is shown. Each column represents a different slice. The first (top) row, denoted as row A, shows images showing various slices of the constructed QSM template. The second row, denoted as row B, shows the same image with QSM DPA boundaries for SN 602, RN 604, and STN 606. The third row, denoted as row C, shows the boundaries transformed to the original space for SN 602, RN 604, and STN 606. The final row, denoted as row D, shows images showing various midbrain slices of a subject with the final boundaries of SN 602, RN 604, and STN 606 obtained after applying DPA to the boundaries in row C.

For initial template training, 30 cases were processed for MTC and QSM data. The integrity of the template automated MTC background intensity is demonstrated by the fact that the slope related from one to the other is 0.99, and the R ² is 0.53 and the p-value is less than 0.001. The background value is important for properly thresholding the NM and iron content signals.

Measurements of Dice similarity coefficient and volume ratio (VR) were performed for different thresholds for both MTC and QSM images, but the data associated with NM and SN Dice coefficients plotted against VR were 7A-7B and FIGS. 8A-8B, respectively. Referring to FIGS. 7A and 7B, plots of Dice similarity coefficient plotted against VR values of neuromelanin (NM) are shown based on various scenarios. In both Figures 7A and 7B, plot A corresponds to a scenario in which no threshold is applied, plot B corresponds to a threshold of NM contrast greater than 1000, and plot C corresponds to a threshold of NM contrast greater than 1250. , plot D corresponds to a threshold of NM contrast greater than 1500. The plot in FIG. 7A was generated using data for 30 training cases, while the plot in FIG. 7B was generated using data for 57 validation cases. The average dice, volume ratio, and volume loss associated with template data and manually drawn data are quoted for each threshold. A threshold of 1000 units keeps volume loss to a minimum and produces excellent results.

8A and 8B, plots of Dice similarity coefficient plotted against SN VR values based on various scenarios are shown. In both Figs. 8A and 8B, plot A corresponds to the scenario of no threshold applied, plot B corresponds to a susceptibility value threshold of more than 50 ppb, plot C corresponds to a susceptibility value threshold of more than 75 ppb, and plot D corresponds to a susceptibility value threshold of more than 100 ppb. The plot in Fig. 8A was generated using data for 30 training cases, while the plot in Fig. 8B was generated using data for 57 validation cases. The average Dice, volume ratio, and volume loss associated with the template data and the manually drawn data are quoted for each threshold. A threshold of 50 ppb produces the least loss in SN volume.

The higher the threshold, the denser the distribution becomes, theoretically eventually approaching a single one on both axes. However, higher thresholds also cause higher losses in volume. Therefore, there must be a trade-off between sufficient dice and volume ratio with volume loss. For NM contrast using a threshold of 1000, the average volume loss of template data for all cases is less than 10%, resulting in average DICE and VR values of 0.90 and 0.95, respectively. Higher thresholds, such as 1250, result in an average volume loss of just over 10% (average Dice = 0.92, average VR = 0.96), and data showing NM contrast above 1500 are lower than lower thresholds. also produced an average loss of 20% with relatively high die and VR (0.94 and 0.98, respectively). A threshold of 1000 appears to yield acceptable dice and VR values while keeping volume loss below 10%. Similarly, as shown in Figures 8A-8B for SN, the average susceptibility threshold of 50 ppb resulted in an average volume loss of only less than 10%, while the average Dice and VR values were 0.88 and 0.8%, respectively. It was 94.

FIG. 9 shows a plot of the dice similarity coefficient plotted against the VR value of the red nucleus (RN) based on various scenarios. Plot A (left column) corresponds to a scenario in which no threshold is applied, and plot B (right column) corresponds to a scenario in which a threshold of 50 ppb is applied. The top row of plots has been generated using a dataset of 30 training cases, while the bottom row of plots has been generated using a dataset of 57 validation cases. The average die, volume ratio, and volume loss associated with template data and manually drawn data for the selected threshold are quoted within each plot. A threshold of 50 ppb limits volume loss to approximately 10%.

Figure 10 shows plots of Dice similarity coefficients plotted against VR values for the subthalamic nucleus (STN) under various scenarios. Plot A (in the left column) represents the scenario where no threshold is applied, while plot B (right column) represents the 50 ppb threshold scenario. The top row of plots was generated using a data set of 30 training cases, while the bottom row of plots was generated using a data set of 57 validation cases. The average Dice, volume ratio, and volume loss associated with the template data and the manually drawn data for the selected threshold are quoted within each plot. The 50 ppb threshold keeps the volume loss at approximately 10%.

9 and 10, before applying any threshold to the QSM data, the average dice and volume ratios were 0.93 and 1.06 for RN and 0.76 and 0.95 for STN. However, applying a 50 ppb threshold to the QSM data yields average DICE and VR of 0.95 and 1.04 for RN and 0.83 and 0.98 for STN. An average VR greater than 1 indicates that most of the structures found by the fully automated template/DPA approach tend to be larger than those with the manual/DPA approach.

Referring to FIG. 11, a plot illustrating the correlation between iron content of SN, RN, and STN resulting from manual and template segmentation is shown. The correlation between iron content of SN, RN, and STN resulting from manual segmentation and template segmentation for 30 training cases is shown in the upper plot, while for 57 validation cases. The correlation between the iron content of SN, RN, and STN resulting from manual and template segmentation for is shown in the lower plot. The final results of the template are shown in various plots for ROIs associated with SN, RN, and STN, and manual drawings for iron content using a threshold of 50 ppb. For each of the SN, RN, and STN structures, the slope and ^R2 are shown in each of the plots in FIG. 11. The p-values are less than 0.001 for SN and RN and equal to 0.006 for STN.

Regarding template validation, the validation dataset included 57 cases that were used to process the QSM and MTC data. Dice similarity coefficients were plotted against both NM and SN VR values, and these results are shown in Figures 7B and 8B, respectively. For NM contrast using a threshold of 1000, the average volume loss of template data for all cases was approximately 5%, yielding average DICE and VR values of 0.88 and 1.06, respectively. Similarly, for the average susceptibility of SN, a threshold of 50 ppb produced an average volume loss of approximately 5%, while the average DICE and VR values were 0.89 and 0.97, respectively. Similar to the previous data set, a threshold of 1000 for NM contrast and a threshold of 50 ppb for average sensitivity of SN produced satisfactory results in terms of average DICE, VR, and volume loss across 57 cases.

Dice similarity coefficients for VR for RN and STN are shown in FIGS. 9 and 10, respectively. Before applying any threshold to the QSM data, the average DICE and VR values for RN and STN were 0.92 and 1.05, and 0.74 and 0.86, respectively. Similar to the results from the previous dataset, applying a threshold improved these values. Using a 50 ppb threshold for the QSM data results in an average Dice and VR of 0.95 and 1.03, and 0.81 and 0.90 for RN and STN, respectively, and approximately 12 for both structures. % template average volume loss.

表１．構造ごとの体積推定値、ダイス類似度係数、及び体積比（ＶＲ）の平均値及び標準偏差（ＳＤ）値。 FIG. 11 shows the correlation between iron content of SN, RN, and STN for template data and manual data. For each structure, the corresponding slope and R ² are shown in FIG. The p-values are less than 0.001 for SN and RN and equal to 0.008 for STN. Table 1 below summarizes the results associated with the estimated template volume, VR measure, and Dice similarity coefficient for each structure. The mean and standard deviation values of the first and second data sets and the merged data are shown.

Table 1. Mean and standard deviation (SD) values of volume estimates, dice similarity coefficients, and volume ratios (VR) for each structure.

The studies described with respect to Figures 4-12 used NM and QSM images derived from a single sequence (less than 5 minutes) without the need to manually draw ROIs for non-template-based subjects. , was used for automatic segmentation of SN, STN, and RN in the original space. A multi-contrast atlas combined with DPA for boundary detection both in template space and the original image after transformation back from template space is described and verified. Both Dice values and volume ratios showed excellent agreement for volume and iron content measures between the automated template approach and manual drawing.

Most of the existing templates do not include the NM, SN, STN, and RN and do not include the resolution presented in this work. For example, the MNI template is 1 mm isotropic, while the approach described herein uses 0.67 mm of in-plane isotropic data that is interpolated and then interpolated to 0.167 mm isotropic resolution in 3D. In practice, we found that using whole-brain global deformable registration with 0.67 mm isotropic resolution using either ANT or SpinITK, we were unable to obtain a morphological transformation that could consistently and well match the shape of midbrain structures across subjects. Therefore, to solve this problem and lead to significantly improved results, we added a local registration (also referred to as local transformation here) step. From this local transformation approach, both iron and neuromelanin templates were created from a single multi-echo sequence.

Previous studies have attempted to segment the SN using structural MRI atlases based on T1 and/or T2 weighted sequences. However, the SN is a small structure in the midbrain with low contrast in these images, making it difficult to accurately define the SN boundary. To overcome this limitation, several other studies have attempted to obtain SN atlases using either NM-MRI or QSM. By using a dynamic atlas constructed from NM-enhanced brain images for automatic segmentation of the SN, one group showed a DICE value of less than 0.75. The NM-sensitive T1-weighted fast spin-echo sequence applied in their study is not optimal for clearly delineating the SN. NM contrast can be significantly improved by using MT-MRI acquisition sequences. To the inventors' knowledge, no studies have attempted to automatically segment SN, SNpr, and SNpc using both NM-MRI and QSM.

Previous studies have segmented the SN by including either younger (>5 years and <18 years) or older (>60 years) subjects. The use of atlases based only on young healthy subjects may not be appropriate for studying patients with neurodegenerative diseases affecting midbrain structures. These inter- and intra-subject age-related and disease-related morphometric changes are important and can influence the success of using template approaches when localizing SNs. The most appropriate atlas for a given study is one that requires the least amount of global or local distortion from native subject space to atlas space, and therefore (age: 63.4 years) was used to create an atlas to study patients with neurodegenerative diseases. To reduce the bias caused by intersubject variability, some studies have used probabilistic atlases and constructed probabilistic atlases of SNpc based on NMS-MRI sequences to improve diffusion MRI in PD patients. We attempted to investigate ultrastructural abnormalities in SNpc using this method. They applied symmetric diffeomorphic registration to align T1 images and 27HC SNpc masks onto MNI space and thresholded the atlas with a probability of 50%. However, the average Dice coefficient for Atlas versus human raters was less than 0.61. Thresholding can lead to dramatic changes in the depiction of the NM, thus reducing the reproducibility of the atlas.

Regarding the validation stage, the second data set shows that the dice similarity coefficient and VR values are very close to those from the first data set for the structure, which allows the automatic template proposed herein to Consistency of results produced by the processing approach is verified. Also, the fact that these values are very close to unity indicates the superior performance of this approach.

In the approach described herein, DPA (implementing an edge detection algorithm) is employed after creating a probabilistic atlas of SNpc to further improve the validity of the NM atlas.

In conclusion, the new approach described herein for fully automatic segmentation of deep gray matter in the midbrain (or any anatomical structure in general) uses a template approach as a guide, but dynamically Use with programming algorithms to fine-tune boundaries. Using this approach, it is possible to quantify neuromelanin in the SN and the volume and iron content for all of the SN, STN, and RN. This approach should allow studying changes in these imaging biomarkers for diverse neurodegenerative diseases without the need for manual tracing of these structures.

Regarding neuromelanin (NM) background measurement, since the background value is important for properly thresholding the NM signal, 30 MTC cases were processed to determine the neuromelanin (NM) background measurement. I got it. Referring to FIG. 12, a plot is shown showing the agreement between manual and template neuromelanin (NM) background measurements for 30 healthy controls (or training subjects). Specifically, the plot illustrates the completeness of the template automatic MTC background intensity measure against manual drawings. The agreement between the manual and template MTC background measures shows a slope of 0.99, an R ² of 0.53, and a p-value of less than 0.001.

Regarding DPA, the final template boundaries of all structures of interest can be determined using a dynamic programming algorithm (DPA) for boundary detection in both template space and original space. Exemplary detailed steps of the DPA algorithm may be as follows.
1. Initial boundaries can be drawn on both NM and QSM template data.
2. A background can then be drawn onto the template image, as shown in FIG. 13. In FIG. 13, both NM 702 and background region 704 are shown.
3. The border and background can then be transformed back to the original space, as shown in FIG. 14. In FIG. 14, NM region 802 and background region 804 are shown after transformation from template space to original space. By using the background value plus 1000, we can refine the DPA bounds to provide faster convergence, and a smaller and safer search radius to prevent the algorithm from leaking outside the original bounds. becomes easier to use.
4. Using the background mean + 4σ (where σ is the standard deviation of the background region of interest, which was found to be approximately 250 units in this study) as the threshold for the MTC data, this threshold The NM-refined boundary can be determined by removing all points lower than . For QSM data, pixels with sensitivity less than zero ppb can be removed before DPA is run for all SN, STN, and RN.
A 5.3x3 Gaussian filter can be used to smooth the initial boundaries.
6. A thinning method can then be used to determine the centerline.
7. Boundary refinement before DPA: Global thresholds may not be valid for the entire structure in the MTC data since there may be some variation in intensity around the structure. Therefore, the method of Otsu [Reference: 1979 IEEE “A Threshold Selection Method from Gray-Level Histograms”, Nobuyuki Otsu] is applied to a filtered local image of a 40×40 pixel square around each single centerline point. can be applied and all points below the Otsu threshold can be removed. If the initial boundary is outside the boundary obtained by the Otsu threshold, the boundary can be modified to the nearest remaining point (referred to as the Otsu boundary). This step helps speed up the convergence of DPA.
8. Implementation of DPA: For each point along the centerline, DPA can be performed with the corresponding search box. To allow for curved shapes, the center of the centerline can be used to determine the initial ray. DPA can then be applied to the next set of rays by shifting one pixel along the centerline associated with the initial boundary. For each DPA iteration, the centerline can be updated. Once the end point is reached, these rays can be swept 180°. The algorithm can then move back along the centerline until it sweeps another 180° and reaches the opposite endpoint. Finally, the center point can return along the center line back to the starting point, closing the boundary. This process can be repeated five times, searching both inside and outside the boundaries to get the best results. For NM, STN, and SN, this step can be followed by 5 more iterations using only inner search.

Although the search radius was limited to 4 pixels both inside and outside the boundaries of the expanded space, the cost function used in this DPA was a derivative that avoided leakage of the structure of interest into neighboring objects. and radius of curvature term. For points outside the boundary, negative derivatives were set to zero when searching outward from the center line. This restriction prevented boundaries from leaking into nearby bright objects. The cost function is given below,
During the ceremony,
and the slope and radius along the mth ray at the rth point are denoted by G(r,m) and R(r,m), respectively. The term G _max represents the maximum derivative inside the image and R _avg is the average radius over the previous three radii. The constant α represents the relative weighting of the derivative and radius terms, which can be set to a value between 0 and 1. The closer the shape of the structure of interest is to a circle, the higher α can be set. If there are sharp edges, these can be smoothed by choosing a large α. Since all values of α between 0.05 and 0.15 played a similar role in finding edges faithfully, the value of α=0.1 was chosen to be conservative.

Referring to FIG. 15, images illustrating boundaries reconstructed with different values of the DPA parameter α are shown. The choice of α can dramatically affect the final DPA bound. The NM boundary in the top left image is obtained using α=0.05. In the top right image, the NM boundary has been determined using α=0.10. In the bottom left image, the NM boundary is obtained using α=0.15. In the bottom right image, the NM boundary is obtained using α=0.20. Note that smaller values of α lead to less smoothing from the radius constraints, while larger values of α lead to more smoothing. The higher the value of α, the rounder the structure and the more it tends to shrink toward a circle. Low values lead to boundaries with very zigzag edges. Finally, candidate points for the new boundary can be selected by maximizing the cost function described above for each ray.

Some exemplary simulations are shown in FIGS. 16-18 for various geometries. Referring to FIG. 16, simulation results are shown for using DPA to identify the boundary of a rectangular region with two different intensities. Image (a) shows a rectangular area with two different intensities. Image (b) shows a rectangular region with a border drawn around the central area, and image (c) shows the found centerline and updated border after applying DPA for five iterations. Further shown. Image (d) except that noise was added with a contrast-to-noise ratio (CNR) of 7:1 based on the difference in signal strength between the central region and the frame region within the second rectangular boundary. It is similar to image (b). Image (e) shows the final boundary found after 5 iterations of DPA matching the correct area of 1400 pixels.

For the simulation corresponding to Figure 16, the central area is set to 100 units, the outer frame area within the second boundary but outside the first boundary is set to 30 units, and the back area outside the second boundary is set to 30 units. Ground was set to zero. Gaussian noise with a mean value of zero and a standard deviation of 10 units was added to the image. A first example of a rectangle with sharply defined corners and 1400 pixels is shown in FIG. Despite the presence of noise, the boundaries were still found perfectly.

FIG. 17 shows simulation results for using DPA to identify the boundary of a crescent-shaped region with two different intensities. Image (a) shows a crescent-shaped region with two different intensities, and image (b) shows that the same region was a border drawn around the central area of the crescent-shaped region. Image (c) shows the crescent-shaped region with the corresponding centerline and updated boundaries after 5 iterations of DPA. Images (d) and (e) show centerline and border updates after 10 and 15 iterations of DPA, respectively. Image (f) shows the centerline and boundaries after only 5 iterations using the adaptive Otsu thresholding approach. Image (g) is similar to the noise-added image (c) with a CNR of 7:1 and image (h) with only 5 times of DPA when using the adaptive Otsu thresholding approach. shows the centerline and boundaries after the iterations of .

The crescent-shaped region in Figure 17 was chosen to mimic the more difficult case of detecting the boundaries of curved objects. Figure 17 shows the effect of running different numbers of iterations. The mean value (image (f)) after running a total of 35 iterations was found to be 990.5 with a standard deviation of 1.5, while the Otsu approach in the presence of an SNR of 10:1 Using , the mean value was 984 and the standard deviation was 2.0.

FIG. 18 shows simulation results for using DPA to identify the boundaries of cashew-shaped regions with two different intensities. Image (a) shows a cashew-shaped region with two different intensities, and image (b) shows a cashew-shaped region with an initial border drawn around the central area. Image (c) shows the cashew-shaped region with the corresponding centerline and border updated after 5 iterations of DPA when using the adaptive Otsu threshold approach. Image (d) is similar to image (b), but with added noise and a CNR of 7:1. The final result of the boundary and centerline after 5 iterations of DPA using the adaptive Otsu thresholding approach and in the presence of noise is shown in image (e).

Finally, we selected the cashew-shaped region in Figure 18 to mimic SN and evaluate the method in the absence of sharp edges. After running the first 35 iterations, the mean value was found to be 1086.5 and standard deviation = 2.5. Using the Otsu approach in the presence of a 10:1 SNR and running 30 more iterations, the mean value was found to be 1097 and the standard deviation to be 0.0.

Those skilled in the art will appreciate that the processes described in this disclosure can be implemented using computer code instructions executable by a processor. Computer code instructions may be stored on non-transitory or tangible computer-readable media, such as memory. The memory may be random access memory (RAM), read only memory (ROM), cache memory, disk memory, any other memory, or any other computer readable medium. The processes described in this disclosure can be implemented by an apparatus that includes at least one processor and/or a memory storing executable code instructions. The code instructions, when executed by at least one processor, can cause performance of any of the processes or operations described in this disclosure. The apparatus may be, for example, an MRI scanner or a computing device.

Claims (20)

A magnetic resonance imaging (MRI) system, the system comprising:
an MRI scanner configured to acquire magnetic resonance (MR) data;
at least one processor;
a memory having computer code instructions stored thereon, the computer code instructions being executed by the at least one processor;
acquiring, via the MRI scanner, a corresponding MR image with contrast illustrating one or more anatomical structures of interest for each training subject of a plurality of training subjects;
generating an initial anatomical template image based on first MR data of a first training subject of the plurality of training subjects, the initial anatomical template image defining a template space; to generate;
For each training subject of the plurality of training subjects other than the first training subject,
applying a global transform to the MR image of the training subject to produce, in the template space, a first morphed version of the MR image representing a first estimate of MR data of the training subject; and applying a local transformation to the first morphed version of the MR image of the training subject to create a second morphed version of the MR image representing, in the template space, a second estimate of the MR data of the training subject; and
averaging the initial anatomical template image with the second morphed version of the MR images of the plurality of training subjects other than the first training subject to generate a final anatomical template image;
depicting the boundaries of the one or more anatomical structures of interest in the final anatomical template;
and identifying the boundaries of the one or more anatomical structures of interest for other non-training subjects using the final anatomical template. . In determining the boundaries of the one or more anatomical structures of interest within the MR image, the at least one processor:
MRI system according to claim 1, configured to fine-tune the boundaries of the one or more anatomical structures of interest in the final anatomical template using a boundary detection algorithm. . In determining the boundaries of the one or more anatomical structures of interest within the MR image, the at least one processor:
acquiring, via the MRI scanner, an MR image of the non-training subject with contrast illustrating the one or more anatomical structures of interest;
applying the global transform to an MR image of the non-training subject to generate a first morphed version of the MR image of the non-training subject;
applying the local transformation to the first morphed version of the MR image of the non-training subject to generate a second morphed version of the MR image of the non-training subject;
Projecting the boundaries from the final anatomical template of the one or more anatomical structures of interest onto the second morphed version of the MR image of the non-training subject to determining an estimate of the boundaries of the one or more structures of interest within the template space;
Applying an inverse transformation to the estimate of the boundaries in the template space of the one or more structures of interest to determine the shape of the one or more anatomical structures of interest in the original MR image of the non-training subject. 3. The MRI system of claim 1, configured to: determine a first estimate of the boundary. the at least one processor,
fine-tuning the first estimate of the boundaries of the one or more anatomical structures of interest in the original MR images of the non-training subject using a boundary detection algorithm; 4. The MRI system of claim 3, configured to generate a second estimate of the boundaries of the one or more anatomical structures of interest in the second MR image. 4. The MRI system of claim 3, wherein the inverse transform includes a concatenation of the inverse of the local transform followed by the inverse of the global transform. 2. Applying the local transformation to the first morphed version of the MR image comprises applying the local transformation to a cropping region of the first morphed version of the MR image. MRI system. MRI system according to claim 1, wherein the corresponding MR image comprises a quantitative susceptibility map (QSM) or a neuromelanin image. The at least one processor is configured to determine one of the one or more anatomical structures of interest based on the boundaries of the one or more anatomical structures of interest in the original MR image of the non-training subject. 4. The MRI system of claim 3, configured to determine at least one volume. The at least one processor is configured to determine one of the one or more anatomical structures of interest based on the boundaries of the one or more anatomical structures of interest in the original MR image of the non-training subject. 4. The MRI system of claim 3, wherein the MRI system is configured to determine at least one average intensity. The at least one processor is configured to detect at least one of the one or more anatomical structures of interest based on the boundaries of the one or more anatomical structures of interest in the second MR image. 4. The MRI system of claim 3, configured to determine a sum signal. 1. A method comprising:
acquiring, via an MRI scanner, corresponding MR images with contrast illustrating one or more anatomical structures of interest for each of the plurality of training subjects;
generating, by a computing device, an initial anatomical template image based on first MR data of a first training subject of the plurality of training subjects, the initial anatomical template image defining a template space;
For each training subject of the plurality of training subjects other than the first training subject,
applying a global transformation to the MR images of the training subject to generate a first morphed version of the MR images that represents a first estimate of the MR data of the training subject in the template space; and applying a local transformation to the first morphed version of the MR images of the training subject to generate a second morphed version of the MR images that represents a second estimate of the MR data of the training subject in the template space.
averaging the initial anatomical template image with the second morphed versions of the MR images of the plurality of training subjects other than the first training subject to generate a final anatomical template image;
delineating the boundaries of the one or more anatomical structures of interest in the final anatomical template;
and using the final anatomical template to identify the boundaries of the one or more anatomical structures of interest for other untrained subjects. determining the boundaries of the one or more anatomical structures of interest within the MR image;
12. The method of claim 11, comprising fine-tuning the boundaries of the one or more anatomical structures of interest in the final anatomical template using a boundary detection algorithm. Determining the boundaries of the one or more anatomical structures of interest in the MR image includes:
acquiring, via the MRI scanner, MR images of an untrained subject with contrast illustrating the one or more anatomical structures of interest;
applying the global transformation to the MR image of the untrained subject to generate a first morphed version of the MR image of the untrained subject;
applying the local transformation to the first morphed version of the MR image of the untrained subject to generate a second morphed version of the MR image of the untrained subject;
projecting the boundaries from the final anatomical template of the one or more anatomical structures of interest onto the second morphed version of the MR image of the untrained subject to determine an estimate of the boundaries of the one or more anatomical structures of interest in the template space for the untrained subject;
and applying an inverse transform to the estimates of the boundaries in the template space of the one or more structures of interest to determine a first estimate of the boundaries of the one or more anatomical structures of interest in the original MR image of the untrained subject. fine-tuning the first estimate of the boundaries of the one or more anatomical structures of interest in the original MR images of the non-training subject using a boundary detection algorithm; 14. The method of claim 13, further comprising generating a second estimate of the boundaries of the one or more anatomical structures of interest in the second MR image. 14. The method of claim 13, wherein the inverse transform comprises concatenating an inverse of the local transform followed by an inverse of the global transform. The method of claim 11, wherein applying the local transformation to the first morphed version of the MR image includes applying the local transformation to a cropping region of the first morphed version of the MR image. 12. The method of claim 11, wherein the corresponding MR image comprises a quantitative susceptibility map (QSM) or a neuromelanin image. The at least one processor is configured to determine one of the one or more anatomical structures of interest based on the boundaries of the one or more anatomical structures of interest in the original MR image of the non-training subject. 14. The method of claim 13, configured to determine at least one volume. determining the average intensity of at least one of the one or more anatomical structures of interest based on the boundaries of the one or more anatomical structures of interest in the original MR image of the non-training subject; or determining a sum signal of at least one of the one or more anatomical structures of interest based on the boundaries of the one or more anatomical structures of interest in the second MR image. 14. The method of claim 13, further comprising determining. A non-transitory computer readable medium comprising computer code instructions stored on the non-transitory computer readable medium, the computer code instructions, when executed by a processor, causing the processor to:
acquiring, via an MRI scanner, corresponding MR images with contrast illustrating one or more anatomical structures of interest for each of the plurality of training subjects;
generating an initial anatomical template image based on first MR data of a first training subject of the plurality of training subjects, the initial anatomical template image defining a template space;
For each training subject of the plurality of training subjects other than the first training subject,
applying a global transformation to the MR images of the training subject to generate a first morphed version of the MR images that represents a first estimate of the MR data of the training subject in the template space; and applying a local transformation to the first morphed version of the MR images of the training subject to generate a second morphed version of the MR images that represents a second estimate of the MR data of the training subject in the template space.
averaging the initial anatomical template image with the second morphed versions of the MR images of the plurality of training subjects other than the first training subject to generate a final anatomical template image;
delineating the boundaries of the one or more anatomical structures of interest in the final anatomical template;
and identifying the boundaries of the one or more anatomical structures of interest for other, untrained subjects using the final anatomical template.